does the effect of pollution on infant mortality differ ... · limited if there is a non-linear...

DOES THE EFFECT OF POLLUTION ON INFANTMORTALITY DIFFER BETWEEN DEVELOPING AND

DEVELOPED COUNTRIES? EVIDENCE FROM MEXICO CITY

Eva Arceo, Rema Hanna and Paulina Oliva

Much of what we know about the marginal effect of pollution on infant mortality is derived fromdeveloped country data. However, given the lower levels of air pollution in developed countries,these estimates may not be externally valid to the developing country context if there is a non-lineardose relationship between pollution and mortality or if the costs of avoidance behaviour differconsiderably between the two contexts. In this article, we estimate the relationship between pollutionand infant mortality using data from Mexico. Our estimates for PM10 tend to be similar (or evensmaller) than the US estimates, while our findings on CO tend to be larger than those derived fromthe US context.

Pollution is a grave concern in much of the developing world, with levels that are oftenorders of magnitude higher than in developed countries. Using comparable data,Greenstone and Hanna (2011) document air pollution levels that are five to seventimes higher in India and China than in the US. This may translate into many lost lives:the OECD estimates that almost 1.5 million individuals die from exposure toparticulates each year, many more than who die from malaria or unclean water. Withpollution levels predicted to rise, the OECD claims that this figure may exceed3.5 million people per year by 2050, with most of these deaths occurring in rapidlyindustrialising countries, such as India and China (OECD, 2012).

In contrast with these concerns, Mexico, another rapidly industrialising country, hasexperienced important gains in air quality during the last 20 years. Between 1997 and2006, an array of policies aimed at cutting down pollution in Mexico City resulted inpollutant concentration reductions of between 23% (ozone) and 48% (carbonmonoxide). The policies implemented in this period include centralised fuelimprovement, driving bans, more stringent vehicle and industry emission standardsamong others.1 During the same period, the infant mortality rate dropped by 30% andthe neonatal mortality rate dropped by 20%. Whether or not all – or part of – the time

* Corresponding author: Rema Hanna, Harvard, NBER and BREAD, John F. Kennedy School ofGovernment, Mailbox 26, 79 JFK Street, Cambridge MA, 02138, USA. Email: [email protected].

We thank Jon Hill and Katherine Kimble for excellent research assistance. We are grateful to David Card,Janet Currie, Olivier Deschenes, Heather Royer, Heidi Williams, Catherine Wolfram and Lucas Davis forhelpful comments, as well as seminar participants at CEPR Development Meetings Pre-Conference, the AREBerkeley Seminar, the Environment and Human Capital Conference, the UC Santa Barbara Labor Lunch,UC Berkeley Labor Lunch and El Colegio de Mexico Econ Lunch. We also thank the editor and twoanonymous referees for their very helpful comments. We especially thank Marta Vicarelli for providing uswith the temperature data. This project was funded in part by the Harvard Center for Population andDevelopment Studies and UCMexus.

1 So far no single policy has been proved to be responsible for the sharp drop in pollution. In fact, Davis(2008) finds no effect of the driving ban and Oliva (2015) finds strong evidence of corruption in the smogcheck programme that enforces vehicle emission limits.

[ 257 ]

The Economic Journal, 126 (March), 257–280. Doi: 10.1111/ecoj.12273 © 2015 Royal Economic Society. Published by John Wiley & Sons, 9600

Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA.

series relationship between pollution and infant mortality is causal is still an openquestion.

The challenges with uncovering the causal effect of air pollution on health are wellknown in the economics and epidemiological literature. One of the biggest concerns isthat of attributing to pollution the effect of other factors that may be correlated withhealth, such as weather, socio-economic status and changes in economic conditions.Another important challenge is that of attributing to pollution deaths of individualswho would have died within a few days due to other causes (harvesting). Recent studieswithin economics have overcome some of these challenges using rich data from the USand quasi-experimental approaches that are arguably robust to confounding factorsand harvesting. These methodologies fit generally in one of two groups: fixed effects(Currie and Neidell, 2005) and instrumental variables (Chay and Greenstone, 2003;Knittel et al., 2011).

Many of the existing studies for the developing world are in the epidemiologicalliterature. These studies have typically relied on time series variation in air pollutionwhile controlling for temperature (Borja-Aburto et al., 1998; Loomis et al., 1999) andsometimes controlling for time-fixed unobservable socio-economic or sub-regionalcharacteristics (Borja-Aburto et al., 1997; O’Neill et al., 2004). These empiricalstrategies might be subject to omitted variable bias from unobserved shocks that canaffect both pollution and mortality and are very sensitive to measurement error. Inaddition, none of these studies has examined the effect of carbon monoxide on infanthealth. While one can in principle address omitted variable bias using eitherinstrumental variables or fixed effects, these techniques pose additional challengesin developing countries. For example, a common strategy to find a valid instrument isto use a policy or regulation that can arguably generate exogenous variation inpollution. However, despite the fact that the regulations in developing countries oftenlook similar to those in the US, they are often riddled with implementation andenforcement problems, resulting in a weak first stage.2 The remaining approach, fixedeffects, effectively controls for time-invariant unobserved differences across locationsand overall trends (Currie and Neidell, 2005). This type of empirical model ischallenging when using developing country data, as the measurement error that mayarise from using sparser pollution data may be exacerbated by the inclusion of fixedeffects.3

Studies in the economic literature that aim to have a more causal interpretation arescarce and often lack actual pollution data, which makes it complicated to estimate themagnitude of the effect of pollution concentrations on health ( Jayachandran, 2009;Gutierrez, 2010).4 This comes from the fact that the availability and quality of the air

2 For example, Greenstone and Hanna (2011) experience this problem when using environmentalregulations in India as an instrument for pollution and Davis (2010) finds no effect of driving restrictions onair pollution.

3 As Currie and Neidell (2005) discuss, measurement error has also been noted in the US context as well.Schlenker and Walker (2011) and Knittel et al. (2011) find larger impacts of pollution on health when usingan instrumental variables strategy as compared to fixed effects methods using US data, which both claim isconsistent with classical measurement error being exacerbated with the fixed effects methodology.

4 Other studies include Greenstone and Hanna (2011) who estimate the effect of mandated catalyticconverters on infant mortality rates in India, but have a noisy estimate of the policy impact due to limiteddata; and Tanaka (2015), who measures the effect of more stringent environmental regulation in China.

© 2015 Royal Economic Society.

258 TH E E CONOM I C J O U RN A L [ M A R C H

pollution and mortality data are often more limited in developing countries. Quitefrequently, disaggregated data on infant births and deaths are not accurately recordedor computerised. Even when the data are available, the validity of the data may bequestionable as there is substantial selection as to which births and deaths areregistered. Moreover, there are fewer stations systematically measuring pollution levelsin developing countries, and so there is potentially less variation in pollution to exploit.From this standpoint, relying on estimates from developed countries to estimate thecosts of air pollution in developing countries on health might be attractive option.

There are, however, two important reasons why estimates from developed countriesmay have limited external validity to the developing country context. First, they may belimited if there is a non-linear dose–response relationship between pollution andinfant mortality. If we expect, for example that marginal changes in pollution are moredamaging at higher levels of air pollution, using developed country estimates wouldcause us to grossly underestimate the effect in many developing countries. On theother hand, if there is an inflection point which pollution needs to fall beneath beforehealth gains can be realised, using developed country estimates could alternatively leadus to overestimate the effect.

Second, the effect of pollution on health may be highly dependent on behaviour(Zivin and Neidell, 2009; Moretti and Neidell, 2011; Deschenes et al., 2011). Avoidancebehaviour may be costlier in the developing world, given less access to health care andlower quality housing stock, which would imply that a marginal decrease in pollutionmay have a larger overall health impact in the developing world. Alternatively, theeffect could be smaller if, for example individuals have permanently adapted to badpollution by keeping infants indoors or wearing breathing masks regularly.5 Giventhese two potential factors, applying estimates of the marginal effect of pollution thatare derived from the US to developing countries may be highly misleading for policy.

In this study, we aim to address these problems and estimate the impact of pollutionon infant mortality in a developing country context. To do so, we construct weekly,municipality-level measures of pollution and mortality for 48 municipalities acrossMexico City between the years 1997 and 2006. Mexico City is a highly relevant contextin which to study this relationship. On average, it experiences both the high levels ofpollution and mortality that are common in many developing countries. However,given the high variance in pollution levels, the range of pollution also encompasses arange similar to that observed in the US. These two facts will allow us to estimate themarginal effect of pollution at a range that is typical for developing countries and thento compare this estimate to the marginal effect at the ranges used in the previousestimates for the US.

We first employ a fixed effects technique, controlling for time-invariant character-istics of municipalities, bimonthly 9 municipality fixed effects, weather and munici-pality-specific week trends. Using this method, we find a small effect of pollution onmortality. However, as we discuss below, even with fixed effects, there may beremaining endogeneity concerns. Moreover, despite access to very high qualitypollution measures, station coverage is sparse: depending on the pollutant and year,

5 As higher pollution is more visible, avoidance behaviour may be more likely since the costs of learningabout pollution levels may be lower.


2016] P O L LU T I ON AN D I N F A N T MOR T A L I T Y I N M E X I C O 259

our pollution measures are derived from between 10 and 26 stations. Given that fixedeffects models are particularly sensitive to classical measurement error, our estimatesmay be severely biased downward.

Instead, we exploit the meteorological phenomenon of thermal inversions. Aninversion occurs when a mass of hot air gets caught above a mass of cold air, trappingpollutants. Conditional on temperature, inversions themselves do not represent ahealth risk per se other than the accumulation of pollutants. As such, we can use thenumber of inversions in a given week to instrument for pollution levels that week. Wefind that each additional inversion leads to a 5.7% increase in particulate mattermeasuring 10 lm or less (PM10) and a 6.3% increase in carbon monoxide (CO),conditional on municipality fixed effects, bimonthly 9 municipality fixed effects,municipality-specific time trends, polynomials in temperature and weather controls.

With the instrumental variables strategy, we find robust evidence of pollution oninfant mortality. Our estimates imply that 1 lg/m3 increase in 24-hour PM10 results in0.23 weekly infant deaths per 100,000 births. Similarly, 1 ppb increase in the 8-hourmaximum for CO results in 0.0046 weekly deaths per 100,000 births.6 We find nosignificant effect on neonatal (children 28 days and younger) deaths overall. As a testof the causal pathway, we then separate deaths into those that are likely to be pollutionrelated (i.e. respiratory and cardiovascular disease) versus those that are less likely to bepollution related (i.e. digestive, congenital, accidents, homicides etc.). We findstatistically and policy significant effects of pollution on both neonatal and infantdeaths from respiratory and cardiovascular disease. As we would expect if we hadindeed isolated the effect of pollution from other factors (i.e. income, healthpreferences), we find no effect of pollution on deaths from other causes.

Finally, we compare our estimates to those derived in the US setting. Specifically, wecompare our estimates to Chay andGreenstone (2003), Currie andNeidell (2005), Currieet al. (2009) and Knittel et al. (2011).7 We find larger marginal effects of CO on infantmortality than Currie andNeidell and Currie et al.; we also find larger point estimates thatKnittel et al., but they do not observe a significant effect of CO on infant mortality. ForPM10, our results are near identical to Chay and Greenstone’s results, despite the fact thatthe mean level of pollution in their setting is roughly half of that in Mexico City.

The article proceeds as follows. In Section 1, we describe our empirical methods anddata, while we provide our findings in Section 2. Section 3 provides a discussion of ourestimates with those from the US context. Section 4 concludes.

1. Empirical Method, Data and Summary Statistics

In this Section, we first discuss some of the existing empirical methods for estimatingthe relationship between air pollution and infant mortality. We then detail ourempirical strategy in subsection 1.2. Finally, we describe the data that we collected forthis project.

6 As we illustrate below, these results are robust to different definitions of mortality, different ways tocontrol for seasonality, the inclusion of outliers and different weather and temperature controls.

7 Note that other papers in the US context explore the effect of pollution on child and infant health(Lleras-Muney, 2010). We only include papers that study comparable infant mortality outcomes.



1.1. Existing Empirical Methodologies

Our objective is to estimate the relationship between pollution (Pmw) in a municipality(m) in a given week (w) and mortality per 100,000 live births (Ymw), or the parameter b1:

Ymw ¼ b0 þ b1Pmw þ emw (1)

where ɛmw captures all unobserved determinants of mortality. There are many reasonsfor believing the identification assumption, E(Pmw, ɛmw) = 0, does not hold in this case.For example, areas with low levels of pollution may be richer and thus have lower levelsof mortality regardless of pollution. One method to solve the endogeneity problemwould be to estimate a fixed effects model:

Ymw ¼ b0 þ b1Pmw þ am þ rmj þ emw (2)

where am is a set of municipality fixed effects that control for permanent differencesacross municipalities, such as time-invariant socio-economic characteristics. Similarly,rmj is a set of bimonthly 9 municipality fixed effects, which control for commonfactors in a given two month block that could affect both pollution levels and infantmortality within a municipality.8 The fixed effects model represents a substantialimprovement over the standard cross-sectional regression. However, two concernsremain. First, b1 may still be subject to bias if there are unobservable, time-varyingdifferences across municipalities. One way to account for this is to include munici-pality-specific, linear time trends. However, this may not capture sharp or non-monotonic changes in omitted pollution and infant mortality determinants, such asroad improvements that could result in fewer traffic jams and faster access foremergency vehicles, or similarly, protests and demonstrations that results in disruptedtravel patterns. Second, classical measurement error in the pollution variable will biasb̂1 downwards. Fixed effects estimators exacerbate measurement error, biasing b̂1further towards zero. As compared to developed country settings, this may beparticularly problematic in developing countries, where pollution-monitoring stationsare sparse: for example, as we discuss below, we exploit data from 10 to 26 stations.

1.2. Exploiting Thermal Inversions in an Instrumental Variables Framework

We consider an instrumental variables strategy, which is likely to minimise bias fromboth endogeneity and classical measurement error. Specifically, we exploit a meteo-rological phenomenon: the existence of thermal inversions. Inversions are a commonoccurrence in many cities around the world, ranging from Mumbai, Los Angeles,San Paulo, Salt Lake City, Santiago, Vancouver, Prague etc.9 Air temperature in the

8 We experimented with different ways of modelling the fixed effects of time and location. One naturalway would be to include week fixed effects. However, to be consistent with the IV model below, we include thebimonthly 9 municipality fixed effects (note that we drop the first month pair so that it is not co-linear withmunicipality). The results (both in the fixed effects and IV) look almost identical if we just include bimonthlyfixed effects that are not interacted by municipality, so we decided to include the more restrictive set of fixedeffects. Note that the results are also robust to dropping the time trends and instead including a year fixedeffect.

9 The great smog of 1952 in the UK was caused by an inversion episode and was blamed for upwards of12,000 deaths (Bell and Davis, 2001). This incident sparked greater interest in environmental regulation inthe UK.



troposphere usually falls with altitude at about 6.5°C per 1,000 metres. However,sometimes there is a mass of hot air on top of a mass of cold air; this is called a thermalinversion. There are typically three reasons why this can occur: first, radiationinversions are generated on clear nights when the ground and the air in touch with theground are cooled faster than higher air layers. The conditions for radiation inversionsare more frequent in the winter: under clear conditions, the earth’s infrared emissionswarm the higher layers of air. The cold ground temperatures cool causing the air that isclose to the ground to remain at a lower temperature than the air above. Second,inversions by subsidence occur from vertical air movements when a layer of cold airdescends through a layer of hot air. Third, inversions can also be produced when layersof air at different temperatures move horizontally and a layer of cold air developsbelow a layer of hot air ( Jacobson, 2002).

The thermal inversion does not represent a health risk in itself but when it occurs inconjunction with high levels of vehicle and industrial emissions, it may result in thetemporary accumulation of pollutants (Secretar�ıa del Medio Ambiente, 2005).Specifically when emissions are released in the atmosphere, they rise and can gettrapped in the inversion (see Figure 1). As the sun’s energy equates the temperaturesof the cold and hot air masses, the ‘lid’ effect disappears (the inversion ‘pops’) and thepollutants rise again. Inversions may have substantial effects on the concentrationlevels of certain types of pollutants, particularly primary pollutants (CO, particulatematter, NO, NOx and SOx, VOC) that may be released in the morning rush hourswhen the inversions typically occur (Jacobson, 2002). Out of the primary pollutants, wewould, therefore, expect the largest effects for pollutants in which vehicles comprise alarge share of their emissions. For example, in Mexico City, 98% of CO emissions camefrom vehicles in 1998, and therefore, we expect that a large share is released in themorning commute hours. In contrast, we may expect weaker effects for pollutants likeparticulate matter, in which 36% is released by vehicles, or SO2, in which only 21% is.

Inversions may have muted effects on secondary pollutants (O3, NO2, sulphuricacid), which require time to mix from the primary pollutants, and therefore, may onlyappear later in the day when it is likely that the inversions have already ‘popped’(Jacobson, 2002). Moreover, inversions may inhibit the formation of these pollutantsin other ways. For example, in the particular case of O3, given that the chemical

Thermal inversion layer

(a) (b)

Fig. 1. Thermal Inversions. (a) Without Inversions, Pollutants Rise and Disburse. (b) Pollutants areTrapped Beneath the Inversion Layer



reactions that result in O3 require warmth and sunlight, the thick layers of pollutionassociated with thermal inversions may interfere with O3 formation.10

We can, therefore, formally test whether inversions increase the concentrations ofdifferent types of pollutants (3) and, if so, we can use the number of thermal inversionsin a given week (TIw) to instrument for pollution in (4):11

Pmw ¼ p0 þ p1TIw þX

p2mw þ hðWmwÞ þ am þ rmj þ lmw ; (3)

Ymw ¼ b0 þ b1Pmw þX

b2mw þ hðWmwÞ þ am þ rmj þ emw : (4)

Note that TIw varies at the week level and, therefore, week by year fixed effects are notidentified.12 We, therefore, control for municipality-specific week by year trends (w).We also include municipality fixed effects (am) to control for time-invariantcharacteristics across municipalities and bimonthly 9 municipality fixed effects (rmj)to account for seasonal effects within each municipality.13

Importantly, we include a flexible set of controls for temperature and weatherconditions h(Wmw) that includes a fourth polynomial in mean temperature, a thirddegree polynomial in minimum and maximum temperatures during the week, asecond degree polynomial in precipitation, cloud cover and humidity measures.Controlling for temperature is important for the exclusion restriction to hold, sinceinversions have a clear seasonal pattern and temperature may independently affectinfant mortality (Deschenes and Greenstone, 2011).14 Figure 2, panel (a) shows theaverage number of thermal inversions per week for each month of the year (bars), aswell as the average temperatures for each month of the year (spikes) measured by theright axis. As expected, given the conditions necessary for a radiation inversion, a largeshare of the inversions occurs in the winter (November–March). However, inversionsalso occur in months with relatively high temperatures (April, May and October),which will allows us to disentangle the effects of temperature on infant mortality fromthat of air pollution.

Note four additional specification details. First, all regressions are clustered at theweek level, which is the level of variation in our instrument. However, our estimates arerobust to alternative modelling assumptions for the error term; for example, ourreduced form results remain unchanged if we employ Conley standard errors to adjustfor geospatial correlation (online Appendix Table A1) and our IV results look nearlyidentical if we cluster by both week and municipality (online Appendix Table A2).Second, all regressions are weighted by the number of births in the respective cohort

10 Ozone Formation, EPA, http://www.epa.gov/oar/oaqps/gooduphigh/bad.html#6.11 We have experimented with different ways to model the instrument. For example, interacting inversions

with municipality to allow for differential effects across municipality yields very similar results.12 Note that as we exploit week-to-week variation within municipalities in this setting, sorting across

different municipalities due to differential pollution should not be a large concern. Moreover, Hanna andOliva (2015) show that sorting is not a large concern within Mexico City as very few households move acrosscensus blocks, which are an even smaller geographic unit than municipalities.

13 As we illustrate below, our results are robust to different configurations of the control variables, such asomitting controls for minimum and maximum temperatures during the week, omitting municipality-specifictime trends and including different types of seasonal effects.

14 In addition, including precipitation, cloud cover and humidity is also essential as it is possible that aninversion can lead to a thunderstorm if moisture is trapped in the inversion.



4

3

2

1

Ave

rage

num

ber o

f inv

ersi

ons p

er w

eek

0

4

3

2

1

Num

ber o

f inv

ersi

ons

0

Thermal inversions perweek

Mortality in Mexico city

12

14

15

Tem

pera

ture

in °C

16

17

18

13

Janua

ry

Februa

ryMarc

hApri

lMay Jun

e

Month of the year

July

Septem

ber

Octobe

r

Novem

ber

Decembe

r30

35

40

Wee

kly

infa

nt m

orta

lity

45

50

55

August

Janua

ry

Februa

ryMarc

hApri

lMay Jun

e

Month of the year

July

Septem

ber

Octobe

r

Novem

ber

Decembe

r

August

(a)

(b)

Fig. 2. Thermal Inversions, Temperatures and Infant Mortality, by Month of the Year. (a) Inversions andTemperature. (b) Inversions and Infant Mortality

Notes. Panel (a) of this Figure compares the average number of inversions per week (bars) withthe monthly average temperature in Celsius (spikes) for each month of the year. Panel (b)compares the average number of inversions per week (bars) against the infant mortality rate inMexico City (line) for each month of the year.



(online Appendix Table A1 also shows that the results are not sensitive to theseweights). Third, we can also try to disentangle the effects of different pollutants oninfant health by taking advantage of the fact that the inversion effect may vary based onthe geographical features of a location, such as its altitude. Specifically, this will allowus to create multiple instruments for the pollution variables.

Finally, we can also estimate models that control for mortality in the second largestcity of Mexico, Guadalajara, which shares similar weather patterns to Mexico City butdoes not experience inversions. As Figure 3 illustrates, Guadalajara experiences similarseasonal patterns in mortality to Mexico City. This provides us with an additionalmethod to control for seasonal patterns in mortality that may be due to weather orseasons.

1.3. Data

We compiled a comprehensive data set on pollution measures, weather conditions andmortality for Mexico City for the years 1997–2006. Each data source is described indetail below.

4Mortality in Guadalajara

Mortality in Mexico cityThermal inversions per week

3

2

Num

ber o

f inv

ersi

ons

1

0

Month of the year

20

30

40

Wee

kly

infa

nt m

orta

lity50

60

Janua

ry

Februa

ryMarc

hApri

lMay Jun

eJul

y

Septem

ber

Octobe

r

Novem

ber

Decembe

r

August

Fig. 3. Comparing Mexico City and GuadalajaraNotes. This Figure compares the average number of inversions per week (bars) against the infantmortality rate in Mexico City (bold line) and the infant mortality rate in Guadalajara (dashedline) for each month of the year. Guadalajara’s infant mortality rate appears to be lower andnearly constant across the different months of the year, while Mexico’s City infant mortalityappears to have strong seasonal patterns that coincide with thermal inversion patterns. Thermalinversions are absent in Guadalajara.



1.3.1. Neonatal and infant mortalityWe constructed the mortality measures from data from the Ministry of Health(Secretar�ıa de Salud P�ublica). These data include mortality measures for both theMexico City Metropolitan Area (MCMA) and Guadalajara. We utilise two sources. First,we compiled data from death certificates, including information on day of death,gender of the child, municipality of residence, age of the child at death and cause ofdeath. Second, to compute mortality rates, we gained access to the birth certificateregistry, which contains information on date of birth and municipality of residence.

We then computed weekly, municipality-level neonatal mortality rates (those that are28 days of age and younger) and infant mortality rates (those that are one year old andyounger). To do so, for each week-municipality observation, we calculate the numberof births in the last 28 days and in the last year.15 Mortality rates are then calculated bydividing the total number of deaths in each week-municipality by the total number oflive births in the corresponding age group and then multiplying by 100,000. Thus, ourcoefficients can be interpreted as the number of deaths in a week per 100,000 childrenborn alive in the respective age cohort.

1.3.2. PollutionPollution data are notoriously absent in many developing countries. When available,they are often only cross-sectional, or of mixed quality. In this article, we are able totake advantage of a relatively rich, panel data set that is available for Mexico City,namely the Automatic Network of Atmospheric Monitoring (RAMA). Measures areavailable for particulate matter under 10 micrometres (PM10), sulphur dioxide (SO2),carbon monoxide (CO) and ozone (O3). These data are considered to be of highquality and, as Davis (2008) points out, ‘[t]hese measures are widely used in scientificpublications’ (p. 41). However, it is important to note that they are drawn fromrelatively few stations: PM10 is available for 10 stations from 1997 to 1999, and from 16stations starting in 2000, SO2 is drawn from 26 stations, CO is drawn from 24 stationsand O3 is drawn from 21 stations.

From these data, we construct weekly measures of pollution for each of the 56municipalities in Mexico City using the inverse of the distance to nearby stations asweights (see Currie and Neidell, (2005) for description of the methodology). Out ofthe 56 municipalities in Mexico City for which we have infant mortality data, weinclude the 48 that are within 15 kilometres of a station. There is a trade-off betweenconstraining the sample to municipalities that are even closer to at least one station forgreater precision of the pollution measure and increasing the distance cut-off toinclude more municipalities. As shown in online Appendix Table A3, the key infantmortality results are fairly robust to alternative definitions of this cutoff.

We use the hourly measures of pollution to calculate the maximum daily 8-houraverage for CO and average this over the week, the maximum daily 24-hour average forPM10 and average this over the week, and then weekly averages for SO2 and for O3.

16

15 Since 0.03% of the births certificates have missing month and day of birth, we adjust weekly estimates ofbirths by dividing un-dated births equally among all weeks of the year.

16 We use these measures for ease of comparison with Currie and Neidell (2005) and Knittel et al. (2011).However, our results are robust to alternative measures of pollution.



1.3.3. Thermal inversionsThermal inversions are recorded by the Meteorological Unit of the local Ministry ofEnvironment. They conduct screenings almost every day, which consist of measuringhourly temperatures at different altitudes using an aerostatic balloon.17 The existenceof an inversion is determined upon finding non-monotonic temperature gradients.Records are kept on the time and temperature of the inversion rupture, so that onecan also compute the number of hours an inversion lasted, as well as the thickness of itslayer. We aggregate the data to the weekly level by computing the number of thermalinversions in a given week.

1.3.4. Temperature and weatherWe obtained temperature and weather variables as additional controls in ourspecification. Hourly temperature measures are available from 24 stations in theRAMA network, and daily level measures of humidity, precipitation and cloudmeasures are available daily from 219 local weather stations. Using the samemethodology to compute weekly, municipality-level measures of pollution, we use thisinformation to compute the temperature and weather controls.

1.4. Data Description

We describe the data in Table 1. In panels (a) and (b), we provide information onneonatal mortality rates (for children that are 28 days and younger) and infantmortality rates (those that are one year and younger) respectively. In addition to themeans for the weekly measures used in the regression analysis (column (1)), weadditionally include the mean across municipalities in a given year per 100,000 birthsfor ease of comparison with the US Figures (column (4)).18 Over the period of study,yearly mortality rates are more than double in Mexico City than in the US: the neonatalmortality rate is 1,183, while the US rate is 460. Similarly, the infant mortality rate inMexico City was 1,986, while the comparable US Figure was 698.

Note that we also providemortality estimates by cause of death.We can test whether theeffect of pollutiononmortality is drivenbydeaths thatweexpect tobe related topollution.We define this comparison in two ways. First, we can compare all internal deaths with allexternal deaths (i.e. accident, homicides). This is a very strict definition, in that it assumesthat pollution affects all internal deaths, including, for example those from digestivediseases, which may or may not be affected by pollution. Moreover, there are relativelyfewer deaths from external sources (61.37 per 100,000), with relatively less variation, andso we may not capture an effect with the same sample size due to power concerns.

17 Of the 3,652 days within our sample period, we have data on whether an inversion occurred for 95% ofthese days. We drop the weeks in which we are missing inversion data.

18 Note that the weekly measure for infant deaths in column (1) appears smaller than that of neonatalmortality, despite the fact that the latter also includes neonatal deaths. The difference in magnitudes ismainly due to scaling: neonatal mortality rates are computed by dividing the number of deaths occurred in asingle week within the 28-day cohort by the number of live births corresponding to that cohort. Hence, thedenominator for the neonatal mortality figure is necessarily smaller than the denominator for the infantmortality figure. Column (4) shows mortality rates on a yearly basis. This column is computed by multiplyingcolumn (1) by (4) (or 52/13) in the case of neonatal mortality and by 12 in the case of infant mortality.



Therefore, we also compare diseases that are more likely to be directly attributed topollution (i.e. respiratory and cardiovascular disease) versus those that are less likely to bedirectly attributed to pollution (digestive, congenital, accidents, homicides etc.).

In panel (c), we provide means for particulate matter of 10 lm or less (PM10),carbon monoxide (CO), ozone (O3), and sulphur dioxide (SO2), as well as summaryinformation on inversions.19 Despite falling pollution levels in Mexico City, the averageis still quite high. For example, the mean level of PM10 is about 67 lm as compared to39.45 lm observed in California, as documented by Currie and Neidell (2005).Inversions are fairly frequent: on average, there are 1.7 inversions in a municipality-week. Conditional on an inversion occurring that week, there is an average of 2.77inversions in a municipality-week.

Table 1

Sample Statistics

Mean of deaths in aweek/municipalityper 100,000 births

Standarddeviation Observations

Mean ofdeaths in a

year/municipalityper 100,000 births

(1) (2) (3) (4)

Panel (a): Neonatal mortality rates (28 days and younger)All causes 295.83 520.49 24,691 1,183.34Non-external causes 287.73 513.74 24,691 1,150.93External causes 3.27 49.09 24,691 13.07Respiratory causes 7.85 96.17 24,691 31.39Non-respiratory causes 283.15 507.73 24,691 1,132.61

Panel (b): Infant mortality rates (one year and younger)All causes 38.21 56.93 24,691 1,986.82Non-external causes 36.52 55.14 24,691 1,898.85External causes 1.18 9.61 24,691 61.37Respiratory causes 6.91 29.35 24,691 359.14Non-respiratory causes 30.79 46.65 24,691 1,601.08

Panel (c): Pollution and thermal inversionsParticulate matter 24-hour PM10 66.94 23.85 18,017Carbon monoxide 8-hour avg (CO) 2,707.56 797.70 18,167Sulphur dioxide avg (SO2) 13.30 5.21 18,173Ozone avg (O3) 32.33 7.47 18,167Number of inversions in a week 1.68 1.88 18,538Number of inversions in a week,conditional on an inversion

2.77 1.68 11,257

Notes. This Table provides descriptive statistics for the key variables in the regression analysis. Panel (a)provides information on neonatal mortality, while panel (b) provides information on infant mortality. Panel(c) reports information on each pollutant and the thermal inversions. External cause is defined as deathsfrom accidents and homicides; internal cause encompasses all causes not including accidents or homicides.Respiratory causes (RC) includes respiratory and cardiovascular disease, while non-respiratory includesdigestive, congenital, accidents, homicides etc. Mortality data come from death certificates and were providedby Secretar�ıa de Salud P�ublica. Pollution data come from the Sistema de Monitoreo Atmosferico de la Ciudadde Mexico (www.sma.df.gob.mx). Inversion data were provided by the Meteorological Unit at the Secretar�ıade Medio Ambiente.

19 Given the presence of outliers in the data, we trim the top and bottom 1% of values. The primary resultsremain largely unchanged if we do not trim the outliers as shown in columns (5) and (6) of online AppendixTable A6.



Finally, we describe the evolution of pollution and infant mortality over time. InFigure 4, we graph average weekly neonatal (panel (a)) and infant mortality rates (panel(b)) against each of the four air pollutants over time. Mexico City has been successful inreducing pollution city-wide over the 1997–2006 time period, with all four air pollutants

858075

PM10

in µ

g/m

3PM

10 in

µg/

m3

706560

16141210SO

2 in

ppb

SO2 i

n pp

b

8

161412108

80

70

60

75

65

85

1997 1999 2001 2003

340320300280260

Dea

ths p

er 1

00,0

00 b

irths

340320300280260

Dea

ths p

er 1

00,0

00 b

irths

340320300280260

Dea

ths p

er 1

00,0

00 b

irths

340320300280260

Dea

ths p

er 1

00,0

00 b

irths

45

40

35D

eath

s per

100

,000

birt

hs45

40

35

Dea

ths p

er 1

00,0

00 b

irths

45

40

35

Dea

ths p

er 1

00,0

00 b

irths45

40

35

Dea

ths p

er 1

00,0

00 b

irths

3,5003,0002,500

1,5002,000

AmbientPM10

Neonatalmortality

2005

1997 1999 2001 2003AmbientSO2

Neonatalmortality

2005

1997 1999 2001 2003AmbientPM10

Infantmortality

2005

1997 1999 2001 2003AmbientCO

Neonatalmortality

2005

1997 1999 2001 2003AmbientCO

Infantmortality

2005

1997 1999 2001 2003AmbientO3

Neonatalmortality

2005

1997 1999 2001 2003AmbientSO2

Infantmortality

2005 1997 1999 2001 2003AmbientO3

Infantmortality

2005

CO

in p

pb

3,5003,0002,500

1,5002,000C

O in

ppb

383634

3028

32

383634

3028

32

O3 i

n pp

bO

3 in

ppb

(a)

(b)

Fig. 4. Mortality and Pollution Trends in MCMA. (a) Pollution and Neonatal Mortality. (b) Pollution andInfant Mortality

Notes. This Figure plots the average annual pollution concentrations over time for the maximumdaily 24-hour average of PM10, the maximum daily 8-hour average of CO, the averageconcentration of SO2 and the average concentration of O3 (solid lines). It also plots averageweekly neonatal (panel (a)) and infant mortality rates (panel (b)) for Mexico City (dashed lines).



falling sharply over this period (Molina and Molina, 2002). As the Figures illustrate,mortality rates are also falling and, in some cases, the rates closely track pollution changes.

2. Results

2.1. First Stage Estimates

We begin by examining the relationship between the occurrence of an inversion andeach of the four pollutants (PM10, CO, O3 and SO2), which comprises the first stage ofour instrumental variables strategy. In Figure 5, we graph the average pollutant level bynumber of inversions. As the Figure illustrates, we observe a strong, and fairly linear,relationship between the number of inversions in the last week and PM10 and COlevels. In contrast, there does not appear to be an obvious relationship between thenumber of inversions and either O3 or SO2.

We provide the corresponding regression analysis in Table 2. Specifically, we presentcoefficient estimates from (3). As suggested by the Figure, inversions have a large andsignificant effect on PM10 and CO. One additional inversion in the last week results in a3 lg/m3, or 3.4%, increase in PM10 (column (1)). Similarly, one additional inversion

80

60

40

20

0

15

12

6

9

3

0

0 2 6

2.5 4,000

1,000

0

3,000

2,0001.5

Den

sity

0.5

1

0

2

2.5

1.5

Den

sity

8 hr

ave

rage

CO

in p

pb

0.5

1

0

2

2.5

1.5

Den

sity

0.5

1

0

2

2.5

1.5

Den

sity

0.5

1

0

2

8 84Number of thermal inversions per week

PM10 24 hravg concentration

Histogram ofinversions

0 2 64Number of thermal inversions per week

CO 8 hrconcentration



O3 avgconcentration



SO2 avgconcentration


24 h

r ave

rage

PM

10 in

µg/

m3

Ave

rage

SO

2 in

ppb

4035

25

15

5

30

20

10

0

Ave

rage

O3 i

n pp

b

Fig. 5. The Relationship Between the Number of Thermal Inversion Per Week and PollutionNotes. This figure plots average pollutant concentrations (maximum daily 24-hour average ofPM10, the maximum daily 8-hour average of CO, the average concentration of SO2 and theaverage concentration of O3) by the number of thermal inversions per week (thick bars) andfrequency of thermal inversions per week (thin bars). The horizontal line corresponds to theaverage concentration of the corresponding pollutant in the period of study.



results in a 170 ppb, or 5.6%, increase in CO (column (2)). The effects on PM10 andCO are both significant at the 1% level and both would pass a weak instruments test.We find less precisely estimated effects of thermal inversions on SO2 and O3 (columns(3) and (4)); with F-statistics of 3.146 and 6.060 respectively neither of which pass aweak instruments test. These overall findings are consistent with the theoreticalpredictions discussed in Section 1 that we would expect the largest effects on PM10 andCO.

2.2. Causal Estimates of Pollution on Infant Mortality and Health

In columns (1) and (2) of Table 3, we provide the coefficient estimates of the effect ofeach pollutant on neonatal and infant mortality respectively from estimating (2) (thefixed effects model). Note that we additionally include municipality-specific weektrends in this model.20 In columns (3) and (4), we provide our IV estimates for theindividual effects of PM10 and CO on mortality; here, we report the coefficientestimates from (4).21 As there is no first stage result for SO2 or O3, we do not estimatean IV estimate for these pollutants.

Table 2

The Effect of Thermal Inversions on Pollution (First Stage)

PM10 CO 8-hour SO2 O3

(1) (2) (3) (4)

Inversions 3.311*** 169.884*** �0.257* 0.424**(0.503) (13.451) (0.145) (0.172)

F test 43.37 159.5 3.146 6.060Mean of outcome variable 57.67 2,707.56 13.30 32.33Two-month of year 9 municipality FE X X X XMunicipality fixed effects X X X XWeather controls X X X XMunicipality-week trends X X X XN 18,017 18,167 18,173 18,167

Notes. This table provides the coefficient estimates of the effect of the number of thermal inversions per weekon pollution concentrations, controlling for two-month-of-the-year by municipality fixed effects, municipalityfixed effects, municipality-specific week trends, a fourth degree polynomial in average temperature duringthe week, a third degree polynomial in maximum and minimum temperatures during the week, a seconddegree polynomial in precipitation and cloud and humidity measures. Standard errors (listed below eachestimate in parenthesis) are clustered at the week level. Statistical significance is denoted by: ***p < 0.01,**p < 0.05, *p < 0.10.

20 As online Appendix Table A4 shows, the results are qualitatively smaller if we drop the municipality-week trends (columns (1) and (2)) and look similar if include fewer temperature controls (columns (3) and(4)). We find much smaller effects with Currie and Neidell (2005)’s specification (columns (5) and (6)).

21 Online Appendix Table A5 explores the robustness of the IV estimates to different control variables. Asshown in columns (1) and (2), the effects on neonatal and infant mortality are slightly larger in magnitude ifwe do not include the municipality-week trends. Relaxing the temperature controls (columns (3) and (4))also leads to slightly larger estimates for infant mortality but the effects are still not significant for neonatalmortality. Including a more flexible time trend (columns (5) and (6)) yields similar results.



Tab

le3

The

Effectof

Pollution

onInfantMortality

Fixed

effects

Instrumen

talvariab

les

IVwithmortalityin

GDL

andMTYas

controls

Neo

natal

Infant

Neo

natal

Infant

Neo

natal

Infant

(1)

(2)

(3)

(4)

(5)

(6)

Particu

late

matter24

-houravg(P

M10)in

µg/

m3

0.35

95*

0.06

49**

*0.76

250.23

13**

*0.68

170.20

33**

(0.206

3)(0.023

4)(0.641

0)(0.082

1)(0.611

0)(0.079

2)Carbonmonoxide8-houravg.

(CO)in

parts

per

billion(p

pb)

0.00

400.00

14**

0.01

620.00

46**

*0.01

420.00

40**

*(0.006

0)(0.000

7)(0.013

5)(0.001

6)(0.012

8)(0.001

5)Su

lphurdioxideavg(SO

2)in

parts

per

billion(p

pb)

�0.815

8�0

.123

1*(0.585

7)(0.074

2)Ozoneavg(O

3)in

parts

per

billion(p

pb)

0.31

740.01

76(0.561

2)(0.066

7)Meanofoutcomevariab

les

295.83

38.21

295.83

38.21

295.83

38.21

Notes.ThisTab

lepresentsfixe

deffectsan

dinstrumen

talvariab

leestimates

oftheeffect

ofpollutiononinfantmortality.Eachco

-efficien

tco

rrespondsto

aseparate

regression.Allspecificationsco

ntrolfor2-month-of-the-year

bymunicipalityfixe

deffects,municipalityfixe

deffects,municipality-specificwee

ktren

dsan

dweather

controls.In

theinstrumen

talvariab

lesestimation,thenumber

ofthermal

inversionsper

weekistheex

cluded

instrumen

t.Weather

controlsareafourthdeg

ree

polynomialin

averagetemperature

duringthewee

k,athird

deg

reepolynomialin

maxim

um

and

minim

um

temperaturesduringthewee

k,aseco

nd

deg

ree

polynomialin

precipitation,an

dcloud

and

humiditymeasures.

Stan

dard

errors

(listed

below

each

estimatein

paren

thesis)areclustered

atthewee

klevel.

Statisticalsign

ificance

isden

otedby:

***p

<0.01

,**

p<0.05

,*p

<0.10

.



Using a fixed effects strategy, we find an effect of CO and PM10 on mortality(columns (1) and (2)). As compared to a pure cross-sectional analysis, the fixed effectsestimates, for the most part, tend to be smaller in magnitude, which is consistent withclassical measurement error.

Instead, we turn to our instrumental variables strategy. Here, we find large effects ofpollution on infant mortality but smaller and non-significant effects on neonatalmortality. In the case of infant mortality a 1 lg/m3 increase in PM10 over the weekleads to 0.23 deaths per 100,000 births, while a 1 ppb increase in CO leads to 0.0046deaths.22 This implies that a 1% increase in PM10 over a year leads to a 0.40% increasein infant mortality, while a 1% increase in CO results in a 0.33% increase. Thesefindings are consistent with the fact that deaths by respiratory diseases, which are likelyto be affected by pollution exposure, are a higher share of total deaths for infants thanfor those under 28 days and may reflect underlying differences in ambient airexposure (i.e. infants may be more likely to be taken outside during the day thannewborns).

Next, we explore whether any residual seasonal variation – that is after controllingfor bimonthly 9 municipality fixed effects, temperature and weather – is driving ourresults. First, we can control for mortality in the second and third largest cities ofMexico, Guadalajara and Monterrey, which share similar weather patterns as MexicoCity but do not experience inversions. Including the mortality rate in Guadalajara andMonterrey as controls (columns (5) and (6) of Table 3) does not qualitatively affectthe results. Second, we can confirm whether the effect of pollution is similar acrossseasons, particularly whether it differs across the winter and summer months. OnlineAppendix Table A6 shows these results.23 Note that we observe a significant first stagefor both the summer and winter months.24 On net, the effects on infant mortality donot appear to be significantly different across the winter and summer months. Thus,taken together, we believe that seasonality does not drive our findings.25

Another concern is that models that estimate the effect of pollution on infant deathwithin a short time frame, such as week, overstate the effect. This bias would occur ifpollution simply accelerates infant deaths by a short time period rather than causingadditional deaths (‘harvesting’). We do not find evidence of this. We can see this in twoways. First, the effect on neonatal deaths is much smaller in magnitude than infant

22 Note that Table 3 reports effects of CO in terms of neonatal/infant deaths associated with one part perbillion, which corresponds to an increase in CO concentration of 0.037%.

23 For completeness, we show two possible definitions of seasons. Specifically, we can define summer fromweek 13 to week 42 (columns (1) and (2)) or from week 14 to 43 (columns (3) and (4)). The results arequalitatively similar across the two definitions.

24 The Angrist-Pischke F-statistics are above the Stock-Yogo 10% threshold for weak instruments.25 Online Appendix Table A7 explores the effects of non-linearity. Using information on the mean and

variance from Currie and Neidell (2005), we define a spline at one standard deviation above their mean(39.45 µg/m3 for PM10, and 1998 ppb for CO). This results in two endogenous variables per regression, so weuse indicator variables denoting 1, 2–3 and 4–7 inversions per week as the instrument set. We find suggestiveevidence of non-linearities in the CO effect (Table A7, panel (b)). Importantly, the Angrist-Pischke F-statisticssuggest we can separately identify variation in CO above and beyond the 3,167 ppb threshold by breakingthermal inversions into indicator variables denoting 1, 2–3 and 4–7 inversions per week. We then find thatthe marginal effect of CO is close to zero when CO concentrations are below 3,167 ppb and 0.0092 whenconcentrations are above this threshold (significant at the 5% level). However, we cannot reject that thecoefficients are the same at conventional levels (the p-value for the test of equality of the slopes is 0.256).



deaths. Due to the scaling, we need to multiply the infant mortality estimates tocompare the neonatal and infant mortality estimates. When doing this, it becomesapparent that the effects on neonatal mortality are not only insignificant but also 50%smaller in magnitude. The fact that we find smaller effects of air pollution on childrenbelow the age of 28 days suggests that air pollution is not accelerating the death ofalready vulnerable children but is causing deaths of children with otherwise long lifeexpectancies. Second, in online Appendix Tables A8 and A9, we present the results ofour analysis when we aggregate the data to the month level rather than at the weeklevel. If harvesting is an issue of concern with our estimates, we would expect muchsmaller effects when using data aggregated at the month level. However, this is not thecase.26

It is important to note that our IV results are very robust to changes in the modelspecification, providing a high level of confidence in this strategy. In onlineAppendix Table A10, we first test whether our results are driven by changes in thedenominator of the left-hand side variable, that is, the number of births. This wouldbe the case if pollution shocks have an impact on the number of live births in thecurrent week, which are included in the denominator. We perform this check byestimating our IV model with the log of deaths as a dependent variable and the logof births as an additional control variable. The results of this specification, reportedin columns (1) and (2), can be compared with our main results divided bythe average mortality rate. We find that the results of the log specification forinfant mortality are slightly smaller but not significantly different than our mainestimates, and they are still significant at the 5% level (the effects on neonatalremains insignificant). Thus, it is unlikely that the results are driven by changes inbirths.27

In Table 4, we replicate the IV analysis for mortality that results from differentcauses. This can be viewed as a placebo test: if we find that our pollution measure isresulting in deaths that are unlikely to be related to pollution, we would conclude thatour instrument might be directly linked to unobserved socio-demographic determi-nants of mortality. We can define pollutant-related deaths in two ways. First, wecompare deaths from all types of internal sources and those from purely externalsources. We find no observable effect on either internal or external effects for neonatal(columns (1) and (2)). The effect of pollution on infant mortality appears to be drivenby internal deaths (column (6)); there is no observable effect on external deaths forinfants. However, mortality from external sources is rare compared to that frominternal sources, meaning it would be more difficult to detect an effect on externaldeaths if there indeed is one. Moreover, it is possible that external deaths are simply

26 To make the magnitudes of infant mortality coefficients comparable across online Appendix Table A9and our main results (Table 3), it is useful to compute the yearly mortality effects associated with both (i.e.multiplying coefficients in Table 3 (column (2)) by 52 and the coefficients in online Appendix Table A9(column (2)) by 12). The coefficients from monthly aggregated data appear about three times as large as thecoefficients from weekly aggregated data. Note, however, that only the coefficients from the CO IV regressionare reliable since the first stage for PM10 is lost when aggregating data at such a coarse level (see onlineAppendix Table A8).

27 Note that in columns (3) and (4) of online Appendix Table A12, we also show the estimates had we notdropped the top and bottom 1% of values in pollution and show that the results are not sensitive to theirinclusion.



Tab

le4

IVEffectof

Pollution

onMortality,by

Cau

seof

Death

Neo

natal

Infant

External

Internal

Non-RC

RC

External

Internal

Non-RC

RC

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Particu

late

matter

24-houravg(P

M10)

�0.028

30.42

63�0

.044

80.44

28***

0.01

050.19

02***

0.07

730.12

35***

(0.054

1)(0.474

2)(0.452

1)(0.096

7)(0.008

8)(0.070

1)(0.048

6)(0.034

1)Carbonmonoxide

8-houravg(C

O)

�0.000

80.00

85�0

.001

60.00

93***

0.00

020.00

37***

0.00

140.00

25***

(0.001

1)(0.010

2)(0.009

6)(0.002

1)(0.000

2)(0.001

4)(0.001

0)(0.000

7)Meanofoutcomevariab

le3.27

287.73

283.15

7.85

1.18

36.52

30.79

6.91

Notes.This

Tab

lepresents

theco

efficien

testimates

from

theIV

estimationofeach

pollutantonmortality,

bycause

ofdeath.Eachco

efficien

tco

rrespondsto

aseparateregression.External

cause

isdefi

ned

asdeathsfrom

acciden

tsan

dhomicides;internal

cause

enco

mpassesallcausesnotincludingacciden

tsorhomicides.

Respiratory

causes(R

C)includerespiratory

andcardiovasculardisease,whilenon-respiratory

(Non-RC)includes

digestive,co

nge

nital,acciden

ts,homicides

etc.All

regressionsco

ntrolfortwo-m

onth-of-the-year

bymunicipalityfixe

deffects,

municipalityfixe

deffects,

municipality-weektren

ds,

afourth

deg

reepolynomialin

averagetemperature

duringthewee

k,athird

deg

reepolynomialin

maxim

um

and

minim

um

temperaturesduringthewee

k,aseco

nd

deg

reepolynomialin

precipitationan

dcloudan

dhumiditymeasures.Thenumber

ofthermalinversionsper

wee

kistheex

cluded

instrumen

t.Stan

darderrors(listedbeloweach

estimate

inparen

thesis)areclustered

atthewee

klevel.Statisticalsign

ificance

isden

otedby:

***p

<0.01

,**

p<0.05

,*p

<0.10

.



accidental and/or immediate, and thus are altogether uncorrelated with income orhealth care quality.

Therefore, we can also classify deaths by those that are more likely to beattributed to pollution (i.e. respiratory and cardiovascular disease) versus those fromsources that are less likely (digestive, congenital, accidents, homicides etc.). This isnot a perfect separation, as children who are weakened by high pollution may bemore likely to pass away from other sources (such as digestive disorders) and causesof death may be imperfectly diagnosed. However, we should still expect the effecton respiratory diseases to be relatively large if pollution is driving much of theeffect, and not other socio-demographic characteristics. Our results suggest thatmost of the deaths related to pollution are linked to respiratory and cardiovascularcauses: we find no effect on non-respiratory deaths for either neonatal (column(3)) or infant (column (7)) deaths. Importantly, if we focus on deaths fromrespiratory and cardiovascular disease, we find large effects of pollution on bothinfant (column (8)) and neonatal (column (4)) deaths. The fact that we find thatthe effect of pollution on mortality is driven mainly by respiratory causes impliesthat our instrument is capturing exogenous variation in pollution and not justtrends in socio-economic characteristics.

Finally, in all the IV specifications above, the effect for any individual pollutant maybe capturing its own effect, as well as the effect of the other pollutant affected bythermal inversions (either PM10 or CO). This is particularly problematic given that wehave used one main instrument to identify the effect of each pollutant. In order toaccount properly for the total amount of deaths attributed to pollution fluctuations, weadopt two different approaches. First, we include all pollutants affected by thermalinversions in the same specification, so that the estimated effect of each pollutant ispurged of potential bias from the others. This results in multiple endogenous variablesand, therefore, we need to identify multiple instruments in order to estimate the causaleffect of each pollutant, conditional on the others. Second, we construct a pollutionindex using the principal components method in order to generate a singleendogenous variable that captures information on fluctuations of both PM10 and CO.

To generate multiple instruments, we exploit the fact that the effect of the inversionson different pollutants may differ based on the altitude of the municipality and thethickness of the inversion. Based on this, we create three instruments: the number ofinversions, the number of inversions interacted with altitude and the thickness ofinversion interacted with altitude. As online Appendix Table A11 illustrates, PM10

concentrations are lower at higher altitudes, presumably because higher areas arelikely to be above the thermal inversion layer. However, altitude does not seem tomatter as much for CO concentrations, unless the thermal inversion layer is thick.Thickness of the thermal inversion induces higher concentrations of CO at higheraltitudes. The Angrist-Pischke F-statistic for both estimated equations is above theStock-Yogo 10% critical value for single endogenous regressors, which suggests that ourinstrument combination induces at least some independent variation in bothpollutants.

The IV estimates of this multiple pollutant model are presented in panel (a) ofTable 5; note that given that we have multiple instruments, we estimate the modelusing LIML. We do not observe a significant effect of either pollutant on neonatal



mortality, and the two pollutant variables are jointly insignificant (column (1)). Whileeach pollutant is not an individually significant predictor for infant mortality, the twopollutants are jointly significant in predicting infant mortality at the 10% level (column(2)). In terms of magnitude, we find that given the overall decline in pollution inMexico City from 1997 to 2006 (weighted by births), our estimates would predict achange in the infant mortality rate that is due to pollution of 325 deaths per 100,000(columns (3) and (4)).

Panel (b) of Table 5 shows the IV results where a single pollution index appears asthe unique endogenous variable. The units of the index can be interpreted in terms ofstandard deviations. For reference, the average value of the index is 1.24 in 1997 and�0.68 in 2006. Note that when we collapse the two pollutants into a single variable, theeffect of pollution is statistically significant. This is not surprising given that the indexovercomes the identification problem posed by the high collinearity between PM10 andCO (the correlation is 0.48). Note, that the drop in pollution measured by the units ofthe index results in a very similar reduction in deaths per year as the reduction impliedby the separate measures of the two pollutants: 308 compared to 325 deaths per100,000.

Table 5

Multiple Pollutant Models

Neonatal Infant

Change inpollution between1997 and 2006

Change in infantdeath rate per year

between 1997 and 2006

(1) (2) (3) (4)

Panel (a): Multipollutant modelParticulate matter 24-houravg (PM10)

�0.0598 0.1335 �21.87 �151.8(1.9501) (0.2176)

Carbon monoxide 8-houravg (CO)

0.0185 0.0021 �1,581.73 �172.7(0.0410) (0.0046)

Chi-squared stat jointsignificance test

1.682 9.834 �324.5

p-value 0.431 0.00732

Panel (b): Single pollution indexPollution index 12.5322 3.4382*** �1.72 �307.8

(9.8438) (1.1470)

Notes. In panel (a), each column in this Table presents the results of a single specification that includes COand PM10 simultaneously. We present the results of an IV estimation of the multivariate pollution model,where inversions, inversions 9 altitude and inversions 9 thickness are instruments for CO and PM10 (thefirst stage is reported in online Appendix Table A10). Coefficients can be interpreted as the effect of eachpollutant, conditional on the other pollutant. In panel (b), we use the principal components method toconstruct an index of pollution. This index summarises information on both PM10 and CO. Because thismodel has a single endogenous variable, we use the number of thermal inversions per week as a singleinstrument. All regressions control for two-month-of-the-year by municipality fixed effects, municipality fixedeffects, municipality-week trends, a fourth degree polynomial in average temperature during the week, athird degree polynomial in maximum and minimum temperatures during the week, a second degreepolynomial in precipitation and cloud and humidity measures. Statistical significance is denoted by:***p < 0.01, **p < 0.05, *p < 0.10. For ease of interpretation, in columns (3) and (4), we calculate thechange in infant deaths per year that can be attributed to the change in pollution in Mexico City between1997 and 2006 (assuming 282,000 births per year).



3. Discussion

One of the main goals of this study is to understand better whether pollution estimatesderived from the US context are externally valid to the developing world. If we believethat there is a non-linearity in the relationship between pollution and infant mortality,or that the costs of avoidance behaviour differ between the two settings then estimatesderived from US settings may not be valid in conducting cost–benefit analysis ofenvironmental regulations in the developing world. As we discussed earlier, thedirection of the bias is ambiguous and, therefore, it is hard to benchmark whether wewould be over or under-estimating the benefits.

Thus, we compare our estimates to those from Chay and Greenstone (2003), Currieand Neidell (2005), Currie et al. (2009) and Knittel et al. (2011) in Table 6. Panel (a)reports our estimates, while panel (b) provides the comparable results for the papers inthe US setting. For ease of interpretation, we provide the mean level of infant mortality(column (1)), the mean value of each pollutant (columns (2) and (5)), the pointestimates (columns (3) and (6)) and the elasticity (columns (4) and (7)). Note severalfeatures regarding the Table. First, we use the estimates from the single pollutantmodelsbecause our subset of pollutants differs from these papers. Even though the models arenot fully comparable, we replicate this Table using the multiple pollutant models inonline Appendix Table A12 for completeness. Second, note that Chay and Greenstonestudy total suspended particulates (TSP) and not PM10. For comparability to ourestimates, we follow Knittel et al. and convert the TSP estimates using the followingformula: PM10 = 0.55 TSP. Third, Chay and Greenstone and Knittel et al. study internaldeaths rather than all deaths, so we additionally report the estimates for internal deathsfor our sample. Finally, we put all estimates at the year level for ease of comparison.

Table 6

Comparison with Literature in US Setting (Sample for One Year Olds)

CO Particulate matter

Infantmortality

rateMeanlevel

Effectsize (year) Elasticity

Meanlevel

Effectsize (year) Elasticity

(1) (2) (3) (4) (5) (6) (7)

Panel (a): Mexico City dataInfant mortality 1,986.82 2.71 239.2*** 0.326 66.94 12.3188*** 0.405Infant mortality-internal 1,898.85 2.71 192.4*** 0.274 66.94 9.2196*** 0.348

Panel (b): Estimates from US settingCurrie and Neidell (2005) 391 2.00 16.501** 0.084 39.45 0.013 0.001Currie et al. (2009) 688 1.58 17.6*** 0.040 29.60 �0.189 �0.008Chay and Greenstone (2003) 1,179 35.33 9.47** 0.284Knittel et al. (2011) 280 1.01 40.56 0.146 28.94 17.68** 1.827

Notes. In this Table, we compare our primary estimates with estimates derived from the US context. Weconvert Chay and Greenstone (2003) from total suspended particulars to particulate matter for ease ofcomparison. We also provide the elasticity of the infant mortality rate to pollution (columns (4) and (7)).Given that Currie et al. (2009) does not have a directly comparable group, we report their estimates for thezero to two week age group.



Wefind that a 1 ppm increase inCOover a year leads to 239.2 infant deaths per 100,000births (panel (a)). This implies that a 1% increase in CO over the year leads to a 0.326%increase in the infant mortality rate. We find a much larger effect on the infant mortalityrate than either Currie and Neidell (2005), Currie et al. (2009) or Knittel et al. (2011)(column (3)). The estimated elasticity using theMexicoCity data is larger thanCurrie andNeidell, who find an elasticity of 0.084, or Currie et al., who find an elasticity of 0.04. Ourestimates are larger but not qualitatively different from Knittel et al. (0.146); however, it isimportant to note that Knittel et al. cannot statistically distinguish their estimate from zero.

While we again point out that we are cautious about our non-linear estimation, theysuggest that the effect of particulates is linear. Thus, we expect that our estimates forPM10 should be similar to those from the US. Despite the fact that the overall level ofparticulates is roughly half (66.94 in Mexico City versus 35.33 in the US), both ourpoint estimates for internal deaths and elasticity for PM10 are fairly similar to Chay andGreenstone (2003) (columns (6) and (7)). However, our estimates are much smallerthan Knittel et al., who find that a one unit increase in PM10 in the year leads to 17.68deaths per 100,000 births, or that a 1% increase in PM10 results in a 1.82% increase inthe infant mortality rate.

4. Conclusion

There is a growing concern about the effects of pollutiononhealth in thedevelopingworld.Especially in their urban areas, high population densities and low quality health services arecollidingwithhigh levels ofharmfulpollutant concentrations.This article sheds lighton theimportance of air quality improvements in the effort to curtail mortality rates.

Using a novel instrumental variables strategy, we find statistically significant effects ofpollution on infant mortality in Mexico City. Our estimates imply that a 1 ppb increasein CO over a week leads to a 0.0046 per 100,000 births increase in the infant mortalityrate, while a 1 lg/m3 increase in PM10 leads to a 0.23 per 100,000 births increase intheir mortality rate. This implies that a 1% increase in PM10 over a year leads to a0.40% increase in infant mortality, while a 1% increase in CO results in a 0.33%increase. Our results on CO are generally larger than those estimated with data fromthe US, while we find comparable results for PM10 despite the fact that pollution levelsare more than double in Mexico City. Our findings illustrate that there may bedifferences in estimates for developed and developing countries, suggesting usingestimates from the US setting in benefit–cost calculations may understate the benefitsfrom greater environmental regulation in developing countries.

CIDEHarvard, NBER and BREADUCSB, NBER

Accepted: 7 December 2014

Additional Supporting Information may be found in the online version of this article:

Appendix A. Robustness checks.Data S1.



ReferencesBell, M.L. and Davis, D.L. (2001). ‘Reassessment of the lethal London fog of 1952: novel indicators of acute

and chronic consequences of acute exposure to air pollution’, Environmental Health Perspectives, vol. 109(3), pp. 389–94.

Borja-Aburto, V., Loomis, D., Bangdiwala, S., Shy, C.M. and Rascon-Pacheco, R.A. (1997). ‘Ozone, suspendedparticles, and daily mortality in Mexico City’, American Journal of Epidemiology, vol. 145(3), pp. 258–68.

Borja-Aburto, V., Castillejos, M., Gold, D.R., Bierzwinski, S. and Loomis, D. (1998). ‘Mortality and ambientfine particles in southwest Mexico City, 1993–1995’, Environmental Health Perspectives, vol. 106(12),pp. 849–55.

Chay, K. and Greenstone, M. (2003). ‘The impact of air pollution on infant mortality: evidencefrom geographic variation in pollution shocks induced by a recession’, Quarterly Journal of Economics,vol. 118(3), pp. 1121–67.

Currie, J. and Neidell, M. (2005). ‘Air pollution and infant health: what can we learn from California’s recentexperience?’, Quarterly Journal of Economics, vol. 120(3), pp. 1003–30.

Currie, J., Neidell, M. and Schmieder, J. (2009). ‘Air pollution and infant health: lessons from New Jersey’,Journal of Health Economics, vol. 28(3), pp. 688–703.

Davis, L.W. (2008). ‘The effect of driving restrictions on air quality in Mexico City (2008)’, Journal of PoliticalEconomy, vol. 116(1), pp. 38–81.

Deschenes, O., Greenstone, M. and Shapiro, J. (2011). ‘Defending against environmental insults: drugs,emergencies, mortality and the NOx budget program emissions market’, Working Paper, University ofCalifornia, Santa Barbara.

Deschenes, O. and Greenstone, M. (2011). ‘Climate Change, Mortality, and Adaptation: Evidence fromAnnual Fluctuations in Weather in the US’ American Economic Journal: Applied Economics, vol. 3(4), pp.152–85.

Greenstone, M. and Hanna, R. (2011). ‘Environmental regulations, air and water pollution and infantmortality in India’, Working Paper No. 17210, NBER.

Gutierrez, E. (2010). ‘Using satellite imagery to measure the relationship between air quality and infantmortality: an empirical study for Mexico’, Population and Environment, vol. 31(4), pp. 203–22.

Hanna, R. and Oliva, P. (2015). ‘The effect of pollution on labor supply: evidence from a natural experimentin Mexico city’, Journal of Public Economics, vol. 122, pp. 68–79.

Jacobson, M. (2002). Atmospheric Pollution. History, Science, and Regulation, Cambridge: Cambridge UniversityPress.

Jayachandran, S. (2009). ‘Air quality and early-life mortality during Indonesia’s massive wildfires in 1997’,Journal of Human Resources, vol. 44(4), pp. 916–54.

Knittel, C., Miller, D.L. and Sanders, N.J. (2011). ‘Caution, drivers! Children present: traffic, pollution, andinfant health’, Working Paper No. 17222, NBER.

Lleras-Muney, A. (2010). ‘The needs of the army: using compulsory relocation in the military to estimate theeffect of environmental pollutants on children’s health’, Journal of Human Resources, vol. 35(3), pp. 549–90.

Loomis, D., Castillejos, M., Gold, D.R., McDonnell, W. and Borja-Aburto, V.H. (1999). ‘Air pollution andinfant mortality in Mexico City’, Epidemiology, vol. 10(2), pp. 118–23.

Molina, L.T. and Molina, M.J. (2002). Air Quality in the Mexico Megacity: An Integrated Assessment, Dordrecht:Kluwer Academic Publishers.

Moretti, E. and Neidell, M. (2011). ‘Pollution, health, and avoidance behavior: evidence from the ports ofLos Angeles’, Journal of Human Resources, vol. 46(1), pp. 154–75.

OECD (2012). OECD Environmental Outlook to 2050: the Consequences of Inaction, Paris: OECD.Oliva, P. (2015). ‘Environmental regulation and corruption: automobile emissions in Mexico City’, Journal of

Political Economy, vol. 123(3), pp. 686–724.O’Neill, M.S., Loomis, D. and Borja-Aburto, V.H. (2004). ‘Ozone, area social conditions, and mortality in

Mexico City’, Environmental Research, vol. 94(3), pp. 234–42.Schlenker, W. and Walker, W.R. (2011). ‘Airports, air pollution, and contemporaneous health’, Working

Paper No. 17684, NBER.Secretar�ıa del Medio Ambiente (2005). Informe Climatol�ogico Ambiental del Valle de M�exico. Mexico: Gobierno

del Distrito Federal.Tanaka, S. (2015). ‘Environmental regulations on air pollution in China and their impact on infant

mortality’, Journal of Health Economics, vol. 42(July), pp. 90–103.Zivin, J.G. and Neidell, M. (2009). ‘Days of haze: environmental information disclosure and intertemporal

avoidance behavior’, Journal of Environmental Economics and Management, vol. 58(2), pp. 119–28.


280 TH E E CONOM I C J O U RN A L [ M A R C H 2016]

does the effect of pollution on infant mortality differ ... · limited if there is a non-linear...

Documents